Spawns 1 process per GPU
(self)
| 35 | self.device_id = device_id |
| 36 | |
| 37 | def multi_card_run(self): |
| 38 | """ Spawns 1 process per GPU """ |
| 39 | init_logger() |
| 40 | |
| 41 | nb_gpu = self.args.world_size |
| 42 | mp = torch.multiprocessing.get_context('spawn') |
| 43 | |
| 44 | # Create a thread to listen for errors in the child processes. |
| 45 | error_queue = mp.SimpleQueue() |
| 46 | error_handler = ErrorHandler(error_queue) |
| 47 | |
| 48 | # Train with multiprocessing. |
| 49 | process = [] |
| 50 | for i in range(nb_gpu): |
| 51 | self.device_id = i |
| 52 | process.append(mp.Process(target=self.multi_card_train, args=(self.args, self.device_id, error_queue), |
| 53 | daemon=True)) |
| 54 | process[i].start() |
| 55 | logger.info(" Starting process pid: %d " % process[i].pid) |
| 56 | error_handler.add_child(process[i].pid) |
| 57 | for p in process: |
| 58 | p.join() |
| 59 | |
| 60 | def multi_card_train(self, error_queue): |
| 61 | """ run process """ |
no test coverage detected